I'm not sure I understand the problem. If we only need to determine what set produces a sum of square roots that is larger, why can't we simply compare the sum of the original numbers? The square root of 2.0000001, for example, is larger than the square root of 2. The square root of X will be always be larger than the square root of Y if X is larger than Y.
The real problem is calculating the precision of square roots. But the challenge stated in the post can be solved rather easily in a programmatic fashion by simply comparing initial inputs.